Robust chemotaxis beyond sensing limits: signal, noise, and strategy
Robert G. Endres † · 2026
Read Robust chemotaxis beyond sensing limits: signal, noise, and strategy by Robert G. Endres † (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-03632.txt.
Introduction
Robert G. Endres reconsiders a classic puzzle about bacterial chemotaxis, the process by which bacteria such as E. coli swim toward food. Berg and Purcell argued that sensing works near physical limits, yet recent information-theoretic work suggests the cells use only a small fraction of the information in arriving molecules. Endres argues the missing piece is movement strategy.
Using scaling arguments and minimal models, he shows how run-and-tumble motion stays robust to noise through symmetry and time averaging, even when internal information processing is inefficient. Comparing bacterial and eukaryotic chemotaxis sharpens the point. Low information efficiency, he concludes, may reflect an evolved balance of robustness, simplicity, and function rather than poor performance.
Catalog record
| CREATOR | Robert G. Endres † |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 30 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.03632
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Robust chemotaxis beyond sensing limits: signal, noise, and strategy Robert G. Endres † Abstract Bacterial chemotaxis has long been viewed as operating near the physical limits of sensing, as originally articulated by Berg and Purcell. Recent information-theoretic analyses challenge this view, suggesting that Escherichia coli uses only a small fraction of the information available in ligand arrival statistics to bias its motion. How should such low information efficiency be interpreted at the level of behavior? Here, I argue that chemotactic performance is shaped not only by information transmission and noise, but by the strategy of movement itself. Using simple scaling arguments and minimal models, I show how run-and-tumble chemotaxis can remain robust to noise through symmetry and temporal averaging, even when internal information processing is inefficient. Comparing bacterial and eukaryotic chemotaxis highlights how different sensing strategies convert physical limits into observable behavior. These considerations suggest that low information efficiency need not imply poor performance, but may instead reflect an evolved balance between robustness, simplicity, and function. Bacteria navigate chemical gradients by relying on noisy molecular cues, using a biased random walk based on run-and-tumble motion [ 5 , 24 ] . While larger eukaryotic cells perform direct spatial gradient sensing across their cell body, small fast-swimming bacteria perform temporal gradient sensing, that is, by comparing concentration measurements in time [ 17 , 1 , 22 ] . Classical theory has long suggested that this process of chemotaxis operates close to fundamental physical limits of sensing, first articulated for the bacterium Escherichia coli by Berg and Purcell in 1977 [ 4 ] . These limits arise from the stochastic arrival of ligand molecules at the cell surface by diffusion, an unavoidable source of noise that sets a baseline uncertainty for any measurement (see [ 3 ] for a review). Additional fluctuations can be introduced downstream by biochemical signaling networks, which often operate with small numbers of molecules [ 6 ] . Berg and Purcell argued that when framed in terms of signal-to-noise ratios, reliable gradient sensing requires that concentration differences along a ”run” exceed the uncertainty associated with these noise sources [ 4 ] . For physiological parameters and shallow gradients, their analysis indeed suggested that bacterial chemotaxis approaches the predicted limit (see also [ 7 ] for ocean bacteria). Physical sensing limits are not only relevant for bacteria, but also constrain gradient sensing in larger eukaryotic cells, including amoeba, immune cells, and migrating cancer cells [ 15 ] . However, it remains unclear how such physical limits on sensing accuracy translate into effective chemotactic behavior and efficiency over longer time scales, where persistence, memory, and movement strategy play a central role. Recently, Mattingly and colleagues revisited this classic problem, applying an information-theoretic framework to quantify how efficiently E. coli converts stochastic ligand arrivals into directed motion [ 19 ] . Their striking result—that cells use only about one percent of the available information—challenges the long-standing view that bacterial chemotaxis operates near the physical sensing limit. This conclusion is particularly surprising because earlier work using closely related formalisms showed that chemotactic performance is information-limited, with reduced information transmission leading to substantially impaired navigation [ 18 ] . Although phenotypic variability can enhance chemotactic ability in a small subset of cells, the overall finding that typical cells operate far below the information optimum remains unexpected and difficult to interpret. Here, I argue that chemotactic performance is shaped not only by information transmission and noise, but by movement strategy itself. Using simple scaling arguments and minimal models, I show that the run-and-tumble strategy can cancel much of the external and internal noise through symmetry and temporal averaging, even when internal information processing is inefficient. Comparing bacterial and eukaryotic chemotaxis highlights how different sensing strategies convert physical sensing limits into observable behavior. This analysis places the findings of Mattingly et al. in a broader biophysical context. More generally, it suggests that low information efficiency need not imply poor performance, but may instead reflect an evolved balance between robustness, simplicity, and function, where simplicity refers to a minimal control strategy in which cells do not explicitly estimate gradient direction. Information-theoretic limits of chemotactic drift I begin with a brief recap of the information-theoretic framework introduced by Mattingly et al. [ 19 ] . In particular, this framework includes causal versions of the mutual information, measuring correlations between inputs and outputs, as well as rate-distortion theory, which quantifies the minimal information required to achieve biased motion up a gradient for a given acceptable noise level. In particular, they quantify how efficiently E. coli converts external ligand information given by s(t)=d/dt\,\log c=gv_{x}(t) into internal signaling activity, with g=\nabla c/c the relative gradient and v_{x}(t) the effective instantaneous cell swim speed along the gradient in the x direction. The paper defines two dimensionless signal-to-noise ratios (SNRs): one for ligand arrival at the cell surface ( \gamma_{r} ) and another for kinase activity response ( \gamma_{a} ). These are used to compute two corresponding steady-state information rates, \dot{I}_{\dot{s}\to r}^{*} and \dot{I}_{\dot{s}\to a}^{*} , capturing the predictive information from ligand arrival to receptors, and from ligand arrival to downstream biochemical signaling, respectively, using the known physical limit of temporal gradient sensing [ 21 ] . The ratio \eta=\dot{I}_{\dot{s}\to a}^{*}/\dot{I}_{\dot{s}\to r}^{*} quantifies the efficiency of information transmission. However, the measurable quantity is the cells’ drift velocity up the gradient. To ease the discussion and extract intuition, I next introduce a simple information-theoretic toy model connecting ligand arrival statistics (input) to drift behavior (output). Note, this toy model is not intended as a replacement for detailed signaling models, but as a bridge between physical limits and measurable behavior. On the one hand, this provides intuition behind the rigorous and long mathematical derivations in Mattingly et al. ’s papers [ 19 , 18 ] . On the other hand, I would like to make cell-behavioral predictions which can be compared with experimental data. In particular, I am interested in scaling relations. At the most basic level, information about the temporal stimulus arises from estimating changes in concentration during individual runs with speed v . As the cell needs to estimate the stimulus s(t) over a time scale \tau of a run, this results in a concentration change of \Delta\log(c)=gv\tau . The noise in this estimate arises from Poisson fluctuations in ligand arrival, with variance \langle(\delta\log(c))^{2}\rangle\sim 1/(r_{0}\tau) , where r_{0}\sim c is the average arrival rate. Hence, at the input level, the simplified information content about s(t) is [ 12 ] I_{\text{arrival}}\sim\frac{1}{2}\log(1+\text{SNR})\sim\frac{\text{SNR}}{2\ln 2}, (1) with \ln 2 due to conversion from nats to bits. Thus, the signal-to-noise ratio for estimating \Delta\log c is: \text{SNR}\sim\frac{(gv\tau)^{2}}{1/(r_{0}\tau)}=g^{2}v^{2}r_{0}\tau^{3}. (2) Since this estimate is made once every \tau seconds, the information rate becomes: \dot{I}_{\text{arrival}}\approx\frac{\text{SNR}}{2\ln 2\cdot\tau}=\frac{g^{2}v^{2}r_{0}\tau^{2}}{2\ln 2}. (3) Even though the diffusion-limited capture of ligand molecules can be large based on flux J\sim 4\pi Dac\sim r_{0} [ 4 ] , implying relatively small fluctuations in absolute concentration measurements, the relevant signal-to-noise ratio for temporal gradient sensing behaves as \mathrm{SNR}\sim g^{2}r_{0} , which can remain small in shallow gradients. At the output level, we quantify the information in behavior. If the cell does simple random walk e.g. by runs and tumbles, \langle x^{2}\rangle\sim 2D_{\text{eff}}t with D_{\text{eff}}\sim v^{2}\tau . In contrast, if